Multimodal AI · Geodemographics · Remote Sensing

Visual Archetypes of
Indian Geodemographic
Clusters

Using Vision Language Models to extract demographic signals from Sentinel-2 satellite imagery — and testing whether machines can see what surveys measure.

Sentinel-2 · 10m Claude Vision API Google Earth Engine NFHS-5 Alignment Bias-Aware
12
Districts Analysed
4
Clusters
10
Visual Features
694
Parent Dataset
00 · Motivation

Research Questions

India's demographic surveys are rich but infrequent and expensive. Can satellite imagery — analysed by Vision Language Models — provide a continuous, scalable proxy for demographic quality? This study investigates four questions that sit at the intersection of earth observation, large language models, and development policy.

RQ 01
Can VLMs detect visual signatures that meaningfully differentiate geodemographic clusters derived from NFHS-5 survey data?
RQ 02
Do satellite-derived visual feature scores correlate with ground-truth indicators — literacy, stunting, institutional births?
RQ 03
Can LLMs synthesise satellite observations into coherent visual archetype narratives that complement numeric profiling?
RQ 04
What systematic biases emerge when deploying VLMs for geodemographic characterisation at district scale in India?
01 · Methodology

Analysis Pipeline

The pipeline integrates three AI systems — Google Earth Engine for cloud computing over petabytes of satellite imagery, the Copernicus Sentinel-2 constellation for 10-metre resolution optical data, and Anthropic's Claude Vision API for structured feature extraction and narrative synthesis.

NFHS-5 k-means Clusters District Selection (3/cluster) Sentinel-2 via GEE Claude Vision API Feature Extraction NFHS-5 Alignment Archetype Narratives
02 · Input Data

Sentinel-2 Satellite Imagery Grid

True-colour composites (Bands B4, B3, B2) for all 12 districts. Each image covers a 10km × 10km window centred on the district headquarters. 2023 cloud-free annual composite. Rows represent clusters; columns are individual districts.

Satellite Grid

Fig 1 · Sentinel-2 True-Colour Composites · GEE · 2023 · 10m resolution · 10km buffer

03 · VLM Outputs

Visual Feature Heatmap

10 visual features scored 0–10 by Claude Vision API for all 12 districts. Vertical black lines separate clusters. Green = high score, red = low score. Districts within each cluster should show broadly similar patterns if the VLM is discriminating effectively.

Feature Heatmap

Fig 2 · VLM Visual Feature Scores · 12 Districts × 10 Features · Claude Vision API

04 · Cluster Profiles

Radar Feature Profiles

Each cluster rendered as a radar polygon. Non-overlapping polygons indicate effective VLM discrimination. The interactive version is available in outputs/radar_chart.html.

Radar Chart

Fig 3 · Cluster-Level Radar Profiles · Mean VLM Scores · Claude Vision API

05 · Narrative Intelligence

Visual Archetype Narratives

Claude synthesises satellite observations from 3 districts per cluster into a coherent visual archetype — a narrative description of the 'typical' landscape of each geodemographic group as seen from satellite. Top scoring features for each cluster shown as tags.

Cluster 0 — High Deprivation

## Visual Archetype: High Deprivation From satellite, this cluster presents as expansive tracts of dry, brown-to-ochre terrain — laterite plateaus, fallow fields, and seasonal riverbeds — punctuated by sparse forest patches on ridgelines and narrow ribbons of green along floodplains and reservoir margins, with settlements appearing as small, dispersed nucleations rather than consolidated towns. The high land fragmentation (5.0) paired with low road network density (3.3) and near-absent industrial infrastructure (1.3) reflects a landscape where subsistence and rain-fed agriculture dominates, physical connectivity to markets and services remains weak, and the built environment offers few visible markers of economic diversification — conditions that map directly onto chronic poverty, limited institutional reach, and low human development outcomes characteristic of India's most deprived tribal and flood-prone districts. Within-cluster variation is notable: Nuapada presents as rugged, forested hill country with reservoir-dependent agriculture, Pakur as semi-arid rocky terrain along seasonal watercourses, and Purnea as a relatively more urbanized floodplain district with an airstrip and peri-urban sprawl — suggesting that high deprivation manifests through geographically distinct pathways (tribal isolation, resource-poor geology, and flood-vulnerable plains) that nonetheless converge on similarly low levels of visible development.

Agricultural Activity: 5.3 Land Fragmentation: 5.0 Vegetation Cover: 4.7
Cluster 1 — Moderate Deprivation

## Cluster 1: Moderate Deprivation – Visual Archetype From satellite, this cluster typically presents as a small to moderate urban core surrounded by extensive agricultural patchwork in varying stages of cultivation, with fragmented landholdings (land fragmentation: 5.0) and limited industrial infrastructure (1.7), set against a landscape where vegetation cover is moderate but driven more by subsistence farming than planned greenery. The sparse road networks radiating from compact town centers, low urban density, and minimal formal infrastructure visually signal economies tethered to rain-fed or marginally irrigated agriculture, correlating with limited livelihood diversification, constrained service access, and the moderate deprivation outcomes characteristic of India's rural-agrarian heartland districts. Within-cluster variation is substantial: districts range from the green, riverine alluvial plains of Bihar and Uttar Pradesh—where water presence and dense field mosaics suggest productive but poverty-prone floodplain agriculture—to the sandy, near-barren expanses of western Rajasthan, where extreme aridity replaces fragmented fields and deprivation stems less from land pressure than from ecological scarcity, suggesting that distinct environmental pathways converge on similar socioeconomic outcomes.

Agricultural Activity: 5.0 Land Fragmentation: 5.0 Vegetation Cover: 4.3
Cluster 2 — Transitioning

## Cluster 2: Transitioning – Visual Archetype From satellite, these districts present a moderately dense urban core—typically grey-white and sprawling—anchored along a significant water feature (river corridor or lake) that shapes the city's spatial footprint, while the periphery dissolves into semi-arid, brown-toned terrain with fragmented agricultural parcels showing inconsistent cultivation. The moderate urban density (6.3) paired with low green urban space (2.3) and visible peri-urban fringe development suggests rapidly expanding cities where infrastructure and services are stretching to keep pace with growth, likely producing populations with improving but uneven access to healthcare, education, and formal employment—characteristic of districts in active demographic transition rather than established equilibrium. Within-cluster variation is notable: Bhopal's large lake and more radiating road network signal a more planned administrative capital with somewhat higher development, whereas Kurnool's organically grown settlement pattern along a meandering river and drier hinterland suggests a smaller city with weaker peri-urban infrastructure, meaning demographic outcomes likely range meaningfully across this cluster despite shared transitional characteristics.

Urban Density: 6.3 Road Network Density: 6.3 Land Fragmentation: 6.0
Cluster 3 — Better Developed

## Visual Archetype: Better Developed From satellite, these districts present as established urban cores with high-density built-up fabric radiating outward from well-defined city centers, flanked by lateritic or exposed soil on peri-urban fringes where urbanization is actively consuming surrounding land. Prominent water features—whether Kerala's backwater networks, Tamil Nadu's historic tank systems, or Maharashtra's river corridors—thread through the urban mass, while road networks are extensive and industrial zones are clearly delineated, reflecting the economic infrastructure that supports higher employment diversity, stronger service-sector economies, and better health and education outcomes associated with mature secondary cities. Land fragmentation is notably high (6.7), with the dense irregular settlement patterns and visible pockets of informal housing suggesting that rapid in-migration and housing pressure coexist alongside development gains, tempering the uniformly positive demographic profile. Within-cluster variation is significant: Ernakulam's landscape is water-dominated with canal-laced settlement patterns reflecting Kerala's unique high-density-yet-dispersed urbanization model, while Pune and Coimbatore follow a more conventional concentric core-periphery gradient, meaning similar development outcomes emerge from visually distinct spatial configurations.

Urban Density: 7.0 Road Network Density: 6.7 Land Fragmentation: 6.7
06 · Validation

Alignment with NFHS-5 Indicators

Pearson correlation between VLM overall development scores and ground-truth NFHS-5 survey indicators at cluster level (n=4). A strong positive r for literacy and institutional births — and strong negative r for stunting — would validate that satellite visual features carry real demographic signal.

+0.942
Women Literacy
Strong
-0.957
Child Stunting
Strong
+0.963
Inst. Births
Strong
Alignment Scatter

Fig 4 · VLM Score vs NFHS-5 Indicators · Pearson r · 4-Cluster Aggregates

07 · Full Results

District-Level VLM Score Table

DistrictStateCluster DevelopmentUrbanVegetationRoads Land UseConfidence
Pakur Jharkhand High Deprivation 2 2 4 3 mixed 6
Purnea Bihar High Deprivation 4 5 6 5 mixed 7
Nuapada Odisha High Deprivation 2 1 4 2 mixed 6
Sheohar Bihar Moderate Deprivation 3 3 5 4 agricultural 6
Barmer Rajasthan Moderate Deprivation 3 4 1 3 arid 8
Sitapur Uttar Pradesh Moderate Deprivation 4 4 7 4 mixed 7
Nashik Maharashtra Transitioning 6 7 3 7 urban 8
Bhopal Madhya Pradesh Transitioning 7 7 4 7 urban 8
Kurnool Andhra Pradesh Transitioning 5 5 3 5 mixed 7
Ernakulam Kerala Better Developed 6 7 4 6 urban 6
Coimbatore Tamil Nadu Better Developed 7 7 4 7 urban 8
Pune Maharashtra Better Developed 7 7 3 7 urban 8
08 · Responsible AI

Bias & Limitations

Transparent documentation of known methodological limitations is a core contribution of this work. Each issue below represents a direction for future research refinement.

Label LeakageDistrict names provided in prompt. Claude may activate prior knowledge of the area rather than deriving all scores purely from image pixels. A blind-prompt comparison is recommended.
Spatial Scale Mismatch10km buffer around HQ may not represent full district. HQ areas systematically over-represent urban cores, biasing urban_density and road_network_density scores upward.
Seasonal Bias2023 annual composite blends all seasons. Monsoon-driven agricultural patterns (peak June-September) are diluted by dry-season imagery, potentially underestimating agricultural_activity.
VLM Training BiasClaude Vision trained predominantly on Western and Global North imagery. Urban density thresholds calibrated to US/European contexts may not transfer accurately to Indian built environments.
Resolution LimitsSentinel-2 10m resolution cannot resolve individual buildings or narrow lanes. Informal settlement detection at this scale relies on texture patterns, which are unreliable in dense urban areas.
Small n AlignmentCorrelation analysis conducted on n=4 cluster aggregates. With only 4 data points, r values can be driven by single outliers. Results should be treated as directional, not definitive.
Cloud ContaminationResidual cloud pixels in composites — especially in northeastern India — may skew vegetation and brightness scores even after cloud masking.